When you hit send, your message goes to a data centre full of computers. First it's chopped into tokens, about four characters each. Then the model predicts the next token, then the next, one at a time, until your answer is done. It learned that by practising on a huge amount of text, and then people ranked its answers to make it more helpful.
It doesn't learn from your chat while you're talking. Your chats can be used to train future versions, unless you turn off "Improve the model for everyone" in Settings → Data Controls.
All that guessing makes heat, and cooling takes water. How much? Researchers estimated about one 500 ml bottle per 10–50 answers for the older GPT-3 (including water used by power plants). OpenAI's CEO says about one fifteenth of a teaspoon per question for today's model (OpenAI's own estimate). They measure different things, so the honest answer is: it depends who you ask.
Episode 2 of "How it works": everyday tech, zoomed all the way in.
Sources:
OpenAI, how ChatGPT and our foundation models are developed: https://help.openai.com/en/articles/7842364
OpenAI, what are tokens: https://help.openai.com/en/articles/4936856
Ouyang et al. 2022, training language models with human feedback (InstructGPT): https://arxiv.org/abs/2203.02155
OpenAI, Data Controls FAQ: https://help.openai.com/en/articles/7730893
Li et al., "Making AI Less 'Thirsty'": https://arxiv.org/abs/2304.03271
Sam Altman, "The Gentle Singularity": https://blog.samaltman.com/the-gentle-singularity
Which number do you believe? Tell me in the comments. Learned something? A like or a share really helps.
—
I'm Haris. I make videos about the tech that actually matters.
More of my work: https://bytefy.in
#chatgpt #ai #howitworks #techexplained #shorts
